Skip to main content

Grounding \u00b7 Findings

Below 60 pixels, no model reads the sign: find it, crop it, then read it

A sign crop over 100 pixels was readable; under 60 it was not. Finding the sign, cropping it and reading the crop read 7 of 7 speed limits. Asking a 72B vision model to read the whole frame read 1.

Research

In short

  • In street-level images, a sign crop over about 100 pixels was readable and one under about 60 pixels was not, whatever the model. A 101 by 102 pixel speed limit sign was clear; a 57 by 32 pixel give-way sign showed its shape but not its text.
  • Finding the sign, cropping it and reading only the crop read 7 of 7 speed limits with Florence-2-large. Gemini 2.5 Flash, given the whole full-resolution frame, read 6. Qwen2.5-VL-72B and Gemma 3 27B, given frames shrunk to 800 pixels, read 1 and 0.
  • The commonest way to lose a sign is a resize you did not notice.

Speed limits read correctly, 7 signs (share, higher is better)

  1. Florence-2-largefind, crop, read100
  2. Gemini 2.5 Flashwhole frame, full resolution86
  3. Qwen2.5-VL-72Bwhole frame, 800 px14
  4. Gemma 3 27Bwhole frame, 800 px0
7 of 7, 6 of 7, 1 of 7 and 0 of 7. The two open models ran through a hosted API that needed images under about 1 MB, so their frames were shrunk to 800 pixels. April 2026. Source: How we measured.

Why it matters

A traffic sign is often less than 5% of a dashcam frame. The plate under a speed sign that gives the hours is smaller still: about 35 by 20 pixels. A larger model does not help if the pixels never reach it.

How we measured

For the reading comparison we used 7 school-zone speed signs from Berlin with known values, in April 2026. Florence-2-large found the sign, we cropped it with a margin, and read the crop. The other models got the whole frame and the same question. Seven signs is a small set; the gap between 7 and 1 is large, but the size of the frame each model received explains much of it.

The pixel limits come from a later run with 139 sign crops. We measured each crop and recorded whether a vision model could read it.

Where the pixels go

  • A hidden resize. In one pipeline, 101 of 139 crops were shrunk to 512 pixels before the reader saw them. A 35 by 20 pixel plate became about 8 by 5. The reader called 112 of the 139 plates unreadable.
  • The thumbnail, not the original. Mapillary, the public street-level image service, serves a 2,048-pixel thumbnail by default. The original frame is 4,032 by 3,024 pixels or larger. In one Berlin frame a 30 sign is 14 by 15 pixels in the thumbnail and 29 by 30 in the original.
  • Image shape changes cost. Vision models bill images in tiles. For Llama 4 Scout, a tall 149 by 512 crop used 2.2 times the tokens of a 256 by 402 crop with 26% more pixels.

What else we measured

Finding poles was harder than reading signs. Florence-2 alone found poles at an F1 of 0.45, because many poles are small and far away. SegFormer, a segmentation model trained on general scenes, mixed poles with the signs on them, so we dropped it. Masking cars, people and bikes first with a small YOLO model took about 70 milliseconds a frame and was the part of the stack we kept.

What this means if you read text from images

Check the pixel size of the thing you need to read before you choose a model. Get the original image, find the object with a detector or grounding model, crop it, and send only the crop to the reader. If a crop is under 60 pixels, record "not readable" instead of asking. If you need this built for your imagery, tell us about the work.